Profiling transactional applications

Abstract

What does it mean for two transactional applications to be similar? We address this question in this paper by highlighting four distinctive features of transactional applications: (1) the transaction size, i.e., the average number of memory accesses of the transactions; (2) the read-write ratio, i.e., the ratio between the number of accesses that modify the data and those that do not; (3) the contention, i.e., the number of concurrent accesses to the same shared data, such that at least one of these accesses is a write; (4) the uniformity, i.e., the extend to which transactions access distinct objects. We show that the similarity between an application A and an application Aʹ can be derived from these features and can be used to determine which concurrency control implementation works best for A based on having tested which worked best for Aʹ. We convey the accuracy of the profiling and predictions based on a study with six workloads and ten concurrency control mechanisms.